Method and device for adapting a system for object detection

Abstract

The invention relates to a method for adapting a system (S) for object detection, wherein the system (S) has a system architecture consisting of software and hardware components (DK, AI), and at least one of the software components is an AI-based detection component (AI). The method comprises the following steps: a) examining (S2) the system architecture and determining the maximum robustness achievable by the system architecture, b) evaluating (S3) whether the determined robustness meets a predefined threshold, and c) adapting (S4) the system architecture if the determined robustness does not meet the predefined threshold. The invention further relates to a corresponding device (V) for adapting a system (S) for object detection.

Description

[0001] The present invention relates to a method for adapting an object detection system. Furthermore, the present invention relates to a device for adapting an object detection system. Finally, the present invention relates to a computer program product that, on a program-controlled device, initiates the execution of the method for adapting an object detection system.

[0002] Object detection systems are increasingly used in the mobility sector. However, their use in functional safety is limited: IEC 61508-3:2010, Table A.2 "Software design and development - Design of software architecture," states that "artificial intelligence" is "expressly not recommended" for safety integrity levels 2, 3, and 4. Therefore, complex object detection systems have so far only been used to a limited extent in the field of functional safety.

[0003] In non-safety-critical applications, when using artificial intelligence for object detection, the suitability of a system employing AI is first tested. Then, attempts are made to sufficiently demonstrate the desired property of the system using appropriate validation datasets. The problem is that sufficiently comprehensive datasets, especially regarding relevant robustness properties, are difficult or costly to obtain. Therefore, significant effort must be invested, only to potentially discover that the system does not achieve the defined robustness goals. Furthermore, such a result does not indicate whether the goals could be achieved with the chosen system design after improved training or with expanded datasets.

[0004] Regardless of the grammatical gender of a particular term, persons with male, female or other gender identities are included.

[0005] Against this background, one object of the present invention is to provide a way to use a system with at least one artificial intelligence-based component and to ensure that the system meets a certain robustness standard.

[0006] Accordingly, a method for adapting an object detection system is proposed. The system has a system architecture consisting of software and hardware components, where at least one of the software components is an AI-based detection component. The system could, for example, be a mobile system (railway, automotive, etc.) that performs object detection in the vicinity of the vehicle. This could involve, for instance, detecting objects on the track or, particularly in the railway sector, on the rails or in the doorway. The system could also be a general monitoring system, e.g., in the context of industrial automation (where, for example, objects such as items or people within the movement paths of robots could be detected).In general, the system can be used for object detection, especially in safety-critical areas.

[0007] As object detection becomes more complex overall, for example through the use of more sophisticated detection processes such as sensor fusion or through automation, the object detection system comprises a combination of hardware and software components. The hardware components can include detectors (e.g., brightness sensors, temperature sensors, lasers, or similar). These are used in combination with software components that can, for example, process the outputs of the hardware components. At least one of the software components can be an AI-based detection component. The AI-based software component can include a machine learning model or a machine learning algorithm.In particular, the AI-based software component can be used to receive the output of one or more hardware components as inputs and to use them to output a prediction about an object detection or a result of an object detection.

[0008] As explained above, it is particularly important for safety-critical systems to ensure that the necessary quantitative safety objectives can be achieved. This applies especially to the robustness of AI-based detectors, for example, in multi-channel systems, i.e., systems with more than one detector and one detection path. Robustness here refers to the ability of a machine learning model, or the corresponding AI-based detection component, to understand and deliver correct predictions or results, even when the input data is subject to slight changes or disturbances.

[0009] To ensure the fulfillment of quantitative security objectives, the procedure proposed here includes the following steps: a) Examine the system architecture and determine the maximum robustness achievable through the system architecture, b) Evaluate whether the determined robustness meets a predefined threshold, and c) Adapt the system architecture if the determined robustness does not meet the predefined threshold.

[0010] The procedure examines the robustness of the system with its (object) detection function. Robustness always refers to a function provided by the system, which is significantly influenced or defined by the system architecture.

[0011] When examining the system architecture, both individual or multiple detection paths, as well as the entire detection function of the system, can be investigated. Once the maximum robustness achievable by the current system architecture has been determined, the next step is to assess whether this maximum robustness meets a predefined threshold. This predefined threshold can, in particular, represent a robustness level required for commissioning the system and which, for example, corresponds to a safety objective of the system.

[0012] Based on whether or not the threshold is reached, the procedure can then adjust the system architecture or decide that no adjustment is necessary. If the robustness already meets the predefined threshold, the procedure can decide in step c) that no adjustment is required and release the system for operation with the current system architecture. This adjustment step is preferably designed to adapt the system architecture towards the desired robustness. This means that the system architecture is specifically adapted to ensure that the robustness reaches the predefined threshold. As described in more detail below, it is also possible to adjust the system architecture if the desired robustness is even exceeded. This can serve to save costs, for example, by eliminating a redundant path.

[0013] In particular, this method makes it possible to examine a planned system architecture or system design in advance, i.e., before commissioning. This allows for an early assessment of whether a chosen system design is fundamentally capable of meeting the system's robustness requirements. If the targets, i.e., the predefined robustness threshold, cannot be reached, the system architecture can be adjusted in the final step to achieve the required robustness. This step can therefore determine whether the robustness targets are achievable with an adapted system architecture. The method can also be performed during system commissioning, for example, to ensure that the system possesses the required robustness before actual commissioning. Furthermore, the method can be repeated at any time during operation.

[0014] According to one interpretation, the procedure involves repeating steps a) to c) until the predefined threshold is reached. The steps can therefore be performed as often as necessary until an optimum of all objectives is achieved, in particular until the predefined threshold is reached.

[0015] According to another embodiment, the method includes detecting a change in the system architecture and repeating steps a) to c) upon detection of such a change. In this way, the method can also verify, even during system operation, whether the robustness of the system with the modified architecture still meets the predefined threshold.

[0016] According to another embodiment, the method includes defining the system architecture prior to step a). In a preliminary step, the system architecture is defined before determining the maximum achievable robustness. This can be done, for example, using standard solutions. Defining the system architecture also allows for the definition of the resulting fault model (e.g., FTA), which can be used to determine the maximum achievable robustness. If the system architecture has multiple paths, it is assumed that the dependencies or independence of the various detectors / detection paths and their interactions are known, determinable, or estimable.

[0017] Furthermore, defining the system architecture also allows for the setting of a predefined threshold. This can depend, among other things, on the intended use of the system.

[0018] According to another embodiment, determining the maximum achievable robustness includes an empirical evaluation of the system's functionality. Such an empirical evaluation can, for example, involve extrapolation using empirical studies. By examining disturbances and observing the system's response, the robustness of the underlying system architecture can be empirically assessed. Approximating the predefined threshold is possible, for example, by identifying patterns or principles regarding the disturbances and the associated response.

[0019] According to another embodiment, determining robustness includes determining quantitative values ​​of the system's functionality and evaluating these quantitative values. Quantitative values ​​for maximum achievable robustness can be determined, for example, according to "Fawzi, A.; Fawzi, H.; Fawzi, O. Adversarial vulnerability for any classifier. In Proceedings of the NeurIPS 2018, Montreal, QC, Canada, 3-8 December 2018; pp. 1186-1195", https: / / arxiv.org / pdf / 1802.08686.pdf, for the relevant perturbations ("in-distribution").

[0020] For quantitative evaluation, fundamental upper bounds for the robustness of a classifier under perturbation of the input data can be determined according to the reference above. The difference in robustness with respect to unrestricted adversarial examples is then used to quantify the robustness.

[0021] According to another embodiment, adapting the system architecture includes adjusting its complexity. This adjustment can involve both adding one or more additional detection components and / or detection paths to the system architecture (i.e., upgrading the system architecture) and reducing one or more additional detection components and / or detection paths (i.e., downgrading the system architecture).

[0022] If it is determined that the maximum achievable robustness of the system architecture is better than the predefined threshold, i.e., unnecessarily good, the system architecture can be simplified so that the predefined threshold is met, but the maximum achievable robustness is no better than it. In this case, for example, components can be saved by removing detection paths, since the high complexity of the system architecture was not necessary for the objective and a reduced design is sufficient.

[0023] Conversely, the system architecture can be extended with additional detection components or paths to, for example, add further redundancy to the system. This typically increases the robustness of the system architecture. In particular, additional independent detection paths or detectors with a low or well-predicted common cause (CC) fraction can be added.

[0024] According to another embodiment, adjusting the complexity includes reducing the number of classes, reducing the number of inputs to the AI-based detection component, and / or adapting to the specific application of the system.

[0025] Here, the method can improve detection by the AI-based detection component by adapting, and in particular simplifying, the processing by the AI-based detection component. For example, the inputs to the AI-based detection component can be reduced so that it has to process less data and thus becomes more stable in its prediction / result.

[0026] The number of classes used to predict the detected objects can also be reduced. For example, the AI-based detection component can make more general predictions: e.g., object or person instead of which object and which person, small pet instead of hamster, cat, guinea pig, etc. Adaptation to the specific use case or domain-specific adaptation (railway or motor vehicle) is also possible. For a car, a collision with a deer is relevant and the vehicle should brake, whereas for a train, this is not the case. In the case of a train, deer could therefore be categorized as harmless objects that do not trigger a reaction. To achieve this, the training data of the AI-based detection component can be adapted to the specific domain.This also improves the robustness of the system architecture as a whole, as the AI-based detection component can be trained more specifically and therefore becomes better in its predictions.

[0027] Overall, the proposed method provides a way to evaluate an object detection system that includes at least one AI-based detection component. This evaluation is based on a comparison with a decision criterion, a predefined threshold. The corresponding value for the system architecture can be determined, for example, through fault tree analysis or test results. If the robustness of the system architecture meets the target value, i.e., the predefined threshold, the system architecture is acceptable and can be used. If the robustness of the system architecture does not meet the target value, i.e., the predefined threshold, the system architecture can be adjusted accordingly.

[0028] This allows for an early assessment of a system design to determine whether the defined goals are fundamentally achievable. Unnecessary effort and time wasted on further development work can thus be avoided. Furthermore, early adjustments to the system and a re-evaluation of design alternatives are possible. The system architecture, i.e., the functional properties of the system, can also be optimized based on the evaluation performed by this method, given specific robustness goals.

[0029] According to another aspect, a device for adapting an object detection system is proposed, wherein the system has a system architecture consisting of software and hardware components, at least one software component being an AI-based detection component. The device comprises a determination unit, an evaluation unit, and an adaptation unit. The determination unit is configured to examine the system architecture and determine the maximum robustness achievable by the system architecture. The evaluation unit is configured to assess whether the determined robustness meets a predefined threshold. Finally, the adaptation unit is configured to adapt the system architecture if the determined robustness does not meet the predefined threshold.

[0030] The respective unit, for example, the determination unit or the adaptation unit, can be implemented in hardware and / or software. In a hardware implementation, the respective unit can be a device or part of a device, for example, a computer, a microprocessor, or a control unit on a server, a host system, or similar. In a software implementation, the respective unit can be a computer program product, a function, a routine, part of program code, or an executable object.

[0031] The embodiments and features described for the proposed method apply accordingly to the proposed device.

[0032] Furthermore, a computer program product is proposed which, on a program-controlled device, initiates the execution of the procedure described above.

[0033] A computer program product, such as a computer program tool, can be provided or delivered from a server on a network, for example, as a storage medium such as a memory card, USB stick, CD-ROM, DVD, or as a downloadable file. This can be done, for example, in a wireless communication network by transmitting the corresponding file containing the computer program product or tool.

[0034] Other possible implementations of the invention also include combinations of features or embodiments described previously or subsequently with regard to the exemplary embodiments, even if not explicitly mentioned. In such cases, the person skilled in the art will also add individual aspects as improvements or additions to the respective basic form of the invention.

[0035] Further advantageous embodiments and aspects of the invention are the subject of the dependent claims and the exemplary embodiments of the invention described below. The invention will be explained in more detail below with reference to preferred embodiments and the accompanying figures. Fig. 1 shows a schematic block diagram of a device for adapting an object detection system; and Fig. 2 shows a schematic flow diagram of a method for adapting an object detection system.

[0036] In the figures, identical or functionally equivalent elements have been given the same reference symbols, unless otherwise indicated.

[0037] Fig. 1 shows a method for adapting a system S for object detection, as described in Fig. 2 shown.

[0038] System S comprises software and hardware components that together form a system architecture. Fig. 2 Two components, AI and DK, are shown as examples. However, it should be understood that the system S can have any number of components. At least one of the software components is an AI-based detection component, Kl. The system S serves for object detection, for example, in industrial or mobile applications (e.g., railways or motor vehicles). Since it is essential to ensure that Kl components allow for reliable object detection, the system S must be tested before it is actually deployed to determine whether the prediction / detection by the Kl component is sufficiently reliable. For this purpose, the maximum achievable robustness of the functionality of the system S, i.e., its system architecture, is determined and evaluated.

[0039] For this purpose, a device V can be used, as described in Fig. 2 shown, which has a unit of determination BE, a unit of evaluation BWE and a unit of adjustment AE.

[0040] In a first optional step S1 of the procedure of Fig. 1 The system architecture of system S is defined. If the procedure is carried out on an already existing system S, this step can be omitted. Advantageously, however, the procedure can be used to test and potentially adapt the robustness and thus the operational capability of system S during the planning phase.

[0041] The system architecture is then examined (S2). This can be done using the unit of measurement BE. The maximum robustness achievable through the system architecture is also determined in this process.

[0042] In step S3, the evaluation unit BEW assesses whether the specified robustness meets a predefined threshold.

[0043] If this is the case, the system S can be put into operation in step S5.

[0044] If the predefined threshold is not met, the adaptation unit AE in step S4 can adjust the system S or its system architecture. This can involve downgrading the system S if the maximum achievable robustness is unnecessarily high. It can also involve upgrading the system S to adapt the system architecture so that the threshold is met.

[0045] After adjusting the system architecture in step S4, the process starts again with step S2. This can be repeated until the maximum achievable robustness meets the predefined threshold.

[0046] The method and the corresponding device make it possible to examine a system to determine whether a desired robustness is actually achieved, and if not, to adjust the system accordingly. In this way, such a system with the maximum achievable robustness can also be used in safety-critical areas, since it can be ensured in advance that the prediction of the included AI detection components is sufficiently reliable or is compensated for by other components, as the overall robustness of the system meets the specified target value.

[0047] Although the present invention has been described using exemplary embodiments, it can be modified in many ways. Reference symbol list

[0048] AA adaptation unit BE determination unit BWE evaluation unit DK detection component KIKI-based detection component S system V device S1-S5 process steps

Claims

1. A method for adapting a system (S) for object detection, wherein the system (S) has a system architecture consisting of software and hardware components (DK, AI), wherein at least one of the software components is an AI-based detection component (AI), and wherein the method comprises: a) examining (S2) the system architecture and determining the maximum robustness achievable by the system architecture, b) evaluating (S3) whether the determined robustness meets a predefined threshold, and c) adapting (S4) the system architecture if the determined robustness does not meet the predefined threshold.

2. Method according to claim 1, characterized by Repeat steps a) to c) until the predefined threshold is reached.

3. Method according to any one of the preceding claims, characterized by the fact thatThe procedure includes: - Detecting a change in the system architecture of the system (S) and - Repeating steps a) to c) upon detection of a change in the system architecture of the system (S).

4. Method according to any one of the preceding claims, characterized by the fact that The procedure involves defining (S1) the system architecture before step a).

5. Method according to any one of the preceding claims, characterized by the fact that Determining (S2) the maximum achievable robustness includes an empirical evaluation of the functionality of the system (S).

6. Method according to any one of the preceding claims, characterized by the fact that Determining (S2) the robustness includes determining quantitative values ​​of the functionality of the system (S) and evaluating the determined quantitative values.

7. Method according to any of the preceding claims, characterized by the fact that Adapting (S5) the system architecture includes adjusting the complexity of the system architecture.

8. Method according to claim 7, characterized by the fact that Adjusting the complexity involves adding one or more additional detection components (DC, AI) and / or detection paths to the system architecture.

9. Method according to claim 7, characterized by the fact that Adjusting the complexity involves reducing the system architecture by one or more additional detection components (DC, AI) and / or detection paths.

10. Method according to any one of claims 7 to 9, characterized by the fact that Adjusting the complexity includes reducing the number of classes, reducing the number of inputs to the AI-based detection component (AI), and / or adapting to the specific application of the system (S).

11. Device (V) for adapting a system (S) for object detection, wherein the system (S) has a system architecture consisting of software and hardware components (DK, AI), wherein at least one software component is an AI-based detection component (AI), wherein the device (V) comprises: a) a determination unit (BE) for examining the system architecture and determining a maximum robustness of the system architecture achievable by the system architecture, b) an evaluation unit (BWE) for evaluating whether the determined robustness meets a predefined threshold, and c) an adaptation unit (AE) for adapting the system architecture if the determined robustness does not meet the predefined threshold.

12. Computer program product which, on a program-controlled device, causes the execution of the method according to one of claims 1 to 10.

Images